Agent memory means customer conversations, internal documents, and reasoning traces land in a third-party service. That's a data-governance question before it's a technical one.
Enterprise · Security · Governance
A system of record for your agents that your security team can sign off on.
Jennah holds what your agents know, the records they work on, and the decisions people made about them, behind one tenant boundary, whichever models and agent platforms you use. One boundary to secure, one history to audit an agent's memory, and one transaction to erase it.
Jennah is in early access. This page is explicit about what ships today and what does not - including the certifications we don't hold.
The security bottleneck
Why agent state stalls in security review
Developers evaluate agent memory on ergonomics and latency. The people who approve it evaluate something else entirely - and a stack of stitched databases, or a separate memory store inside every agent platform, answers them badly.
When an agent takes a costly action, someone has to reconstruct what it knew at that instant. Logs in one system and embeddings in another can't be read together.
An erasure request has to reach execution logs, embeddings, and graph edges. Spread across three stores, deletion is three scripts, and any one of them can fail after the others have run.
Every agent platform now keeps its own memory. Run agents on two of them, plus a vector and an app database, and each carries its own access model, audit trail, and deletion path. Every one gets reviewed separately.
Production capabilities
What ships today
Available now, running on live infrastructure. Everything in this section is built - the roadmap has its own section below.
Atomic erasure
AvailableDelete an agent's workspace and its execution logs, embeddings, and graph edges are removed together in one transaction, never a partial deletion. The response returns the instant the deletion committed.
History you can audit
AvailableExecution logs are an immutable record of each step an agent took. Facts that change keep their history, so you can ask what was true, or what the agent believed, on the date it acted, not only what the store says now.
One tenant boundary
AvailableEvery read and write is clamped to the enterprise on your credential, server-side, inside the query itself - including graph traversals, which are generated by us rather than accepted as caller-supplied query text. It's an enforced invariant with tests that fail if an unclamped statement can ever be emitted, not an application convention.
Roles and permissions
AvailableRole-based access control with custom roles, so operators, engineers, and auditors get different reach into the same enterprise. Membership is many-to-many: one person can hold different roles in different enterprises.
Scoped, revocable machine credentials
AvailableAPI keys for agents and CI: minted by an admin, shown once, stored only as a hash, with optional expiry and instant revocation. Each key carries its own permissions, fixed at creation (a read-only key is one choice), and can be limited to specific agents or datasets, so a leaked key reaches only what it was minted for.
Region pinning and residency
AvailableEach agent's memory is pinned to a home region and served from the nearest data plane - a single region, one geography, or global. Stored memory and its embeddings stay in that region. Memory formation and optional reranking call a model service that is not region-pinned, unless you bring your own model.
Human approvals with a permanent record
AvailableAn agent asks a named person before it acts. Each decision is append-only and carries who decided, when, and a digest of exactly what they were shown. Approvals can require a signed-in member, or a decider other than the requester. Your agent checks the decision before it proceeds; Jennah records it.
Any model, any agent platform
AvailableAgents reach Jennah over REST, gRPC, and SDKs for Go, Python, and TypeScript, whatever model or framework they run on. Change models or platforms and the state, its history, and its access rules stay where they are.
Bring your own formation model
AvailableOn paid plans, memory formation can run on a model in your own model account, in a region you choose. Conversation content then goes under your model contract instead of ours, and we never fall back to our model if yours fails.
One system of record instead of many
Most teams assemble agent memory from a vector database, a graph database, and a relational store, then write glue to keep them in sync. Jennah keeps memory and application data in one store, with approvals behind the same boundary - which changes the cost structure and the review surface at the same time.
One store to provision, upgrade, back up, monitor, and capacity-plan - instead of three, each with its own failure modes and on-call surface.
One access model, one encryption posture, one set of credentials and network paths. Your security team reviews a single system, once.
Logs, embeddings, and graph edges commit in one transaction and are readable the instant it lands. There's no window where an agent's graph disagrees with its vectors.
We won't quote you a savings percentage from someone else's deployment. The saving here is structural; what it's worth is a figure from your own bill and your own headcount, and we'll help you build it from your workloads.
On the roadmap - talk to us
The standard enterprise checklist, and where we honestly are on each item.
Nothing in this section is available today. It's here so you can see the plan, weigh it against your own timeline before we take up your security team's week, and tell us what to prioritise. Your feedback has real influence over this order.
Dedicated single-tenant instances
DesignedA private database or instance rather than a shared multi-tenant one, with its own region, capacity, lifecycle, and embedding model. Designed and specified; not yet built. Today every tenant shares a multi-tenant database with row-level isolation.
SSO / SAML 2.0
PlannedOkta, Microsoft Entra ID, and other SAML identity providers, with SCIM provisioning to follow. Today sign-in is social OAuth login for people and enterprise-scoped API keys for workloads.
Exportable audit log
PlannedAdministrative and access events exported or streamed to your SIEM. Today an agent's own execution history is queryable through the API, but control-plane audit export isn't built.
Customer-managed keys
PlannedCMEK on the underlying managed database, so key rotation and revocation sit with you. Today data is encrypted at rest and in transit with provider-managed keys.
Compliance program
Not startedWe hold no certifications today - no SOC 2, no ISO 27001, no HIPAA attestation - and we won't imply otherwise with a badge strip. If a specific certification gates your purchase, tell us which one and by when: that's the input that decides what we do first.
Starting a security review
Security review is where enterprise infrastructure deals go to die, usually because nobody said early what the vendor couldn't do. So here's both halves.
What we can put in front of your team this week
- An architecture summary: what runs where, what's stored, what leaves the region
- The tenancy and isolation model, down to how the tenant clamp is enforced and tested
- Encryption at rest and in transit, and our subprocessor list - the cloud provider and the model providers used for embeddings, memory formation, and reranking
- Data handling for embeddings: what content is sent for vectorisation, and what's retained
- A data processing agreement on request
- A direct line to the founder who built the system, not a support queue
What we can't do yet
- Hand you a SOC 2 report, an ISO 27001 certificate, or a HIPAA attestation
- Run the data plane inside your own cloud account
- Give you customer-managed encryption keys or SSO today
- Turn around a 300-question vendor questionnaire in a week - we're a small team and we'll tell you honestly how long it'll take
- Point you at a peer reference in your industry - we're early, and we'd rather say so than imply one
Procurement: a listing that lets you buy through your existing AWS agreement and committed spend is in progress - not live yet. If that's how your organisation prefers to onboard vendors, it's worth telling us; it moves up the queue.
Where this fits best
We're not going to claim customers we don't have. These are the shapes of problem Jennah was designed around - long-lived agent state, a real need to explain past decisions, and data that can't be casually copied between systems.
Financial services
Research, risk, and onboarding agents where every action needs a defensible trail: the steps the agent logged, the facts it held on that date, and who signed off before it acted. That is the difference between an explanation and a guess.
Healthcare and life sciences
Clinical and trial-facing agents that need hard tenant isolation, regional residency, and one place to keep each patient's memory, embeddings included, that a single call erases.
Support and IT automation
Long-running resolution agents that carry customer context across months. Recall the relevant history instead of replaying whole transcripts into the context window on every turn.
Talk to us
Tell us what your security team will ask and what your timeline looks like. You'll get a straight answer about whether we're ready for you yet - and if we aren't, we'll say when we expect to be.
Prefer email? Reach us at info@nightblue.io.